Issue about darken image by Cyclegan and CUT

[pyyyyyyyyy]

Thanks for your great work! I want to know the details about your darken experiment with CUT and Cyclegan, hope your reply!

The WIDER FACE trainset has 12880 images and darkface has 5500 images, I set 12880 widerface to trainA and 5500 darkface to trainB in official Cyclegan model for 200 epoches, but it seems training so many times and got worse result, which i think cannot use to retrain the DSFD model in supervised manner.

So i want to know your training details about the generation of darken images. Thank you, appreciate for you time!

Those are pairs of test images in Widerface i got.

[daooshee]

According to my experience, CycleGAN and CUT are indeed hard to train on WIDER FACE and DARK FACE. My solution is to generate results every once in a while and pick the best checkpoint. However, even this way, the generated results are not satisfactory.

Sorry that since this project is a long time ago, it is hard to find my original settings for this experiment. I remember I did not use resize (which would change the height-width ratio) and used a resolution larger than the original setting in CycleGAN and CUT.